An approximate framework for quantum transport calculation with model order reduction

Quan Chen, Jun Li, Chiyung Yam, Yu Zhang, Ngai Wong, Guanhua ChenJournal of Computational Physics 2015 · OpenReview

Introduction

As semiconductor technology approaches the 10nm node, the design and optimization of transistors require high-fidelity modeling of quantum mechanical effects. The Non-Equilibrium Green’s Function (NEGF) formalism is the standard method for such simulations. However, NEGF presents two significant computational hurdles: the "cross-section bottleneck," where complexity scales cubically with device area, and the "energy integration bottleneck," requiring independent calculations across hundreds of energy points.

This paper presents a novel framework that addresses these challenges by combining a new "X-formulation" for sparse linear systems with adaptive Model Order Reduction (MOR). By exploiting the mathematical correlations between energy states, the proposed method achieves substantial speedups and memory savings without sacrificing engineering accuracy.

The X-Formulation: A State-Space Approach

Traditional NEGF implementations, such as the Recursive Green’s Function (RGF) or Selected Inversion (SelInv), focus on computing specific elements of a matrix inverse. In contrast, the authors introduce the X-formulation, which reformulates the transport problem into a sparse linear system of the form $AX = B$.

A key innovation in this formulation is the use of low-rank approximations for the self-energy matrices, which represent the device's leads or contacts. This transformation allows the system to be viewed in a "state-space" representation. This shift is critical for two reasons:

  1. Scalability: It enables the use of modern iterative solvers, which scale more efficiently than the direct solvers used in RGF for large cross-sections.
  2. Compatibility: It provides the mathematical structure necessary to apply advanced Model Order Reduction techniques originally developed in the fields of systems and control.

Adaptive Model Order Reduction (MOR)

The second pillar of the framework is a nonlinear MOR scheme designed to bypass the energy integration bottleneck. Because physical properties change smoothly across the energy spectrum, the solutions at neighboring energy points are highly correlated.

The authors utilize a Parameter-Dependent Krylov Subspace (PDKS) approach. Instead of solving the full-scale system at every required energy point, the framework solves the "full" problem only at a few strategically chosen interpolation points. These solutions form a basis used to project the massive original system onto a significantly smaller subspace, creating a Reduced Order Model (ROM). Solving the ROM for the remaining energy points is computationally negligible.

To ensure accuracy, the framework employs an adaptive greedy search algorithm. This algorithm automatically identifies the best energy points for full-scale solving by targeting areas where the residual error is highest. Furthermore, the use of tangential interpolation via Singular Value Decomposition (SVD) allows the framework to select only the most significant directions in the right-hand-side vectors, further streamlining the reduction process.

Experimental Evaluation and Results

The framework was benchmarked using three 3D silicon nanowires (SiNW) of increasing size, with the largest (SiNW3) featuring a matrix size of approximately 203,000.

Performance and Speedup

The results demonstrate a dramatic improvement over traditional methods. For the largest device (SiNW3), the X-formulation achieved a 16x speedup compared to the standard RGF algorithm. The MOR scheme proved highly efficient, reducing the number of full-size system solves by a factor of 2.5x to 5x. For instance, in the SiNW1 case, the framework required full solves for only 77 out of 500 energy points to maintain high accuracy.

Memory Efficiency

Memory consumption is a frequent barrier in large-scale quantum simulations. The X-formulation significantly reduced the memory footprint; for the largest benchmark, it required only 7.2 GB of RAM, compared to the 51 GB required by traditional methods. This reduction makes it feasible to perform complex simulations on standard workstations rather than specialized high-performance computing clusters.

Accuracy

Despite being an approximate framework, the method maintains rigorous accuracy. The error was consistently controlled within the $10^{-6}$ to $10^{-7}$ range, which is well within the requirements for semiconductor device engineering.

Conclusion

The proposed framework represents a significant step forward in making quantum transport simulations practical for large-scale device design. By bridging the gap between quantum physics and model order reduction through the X-formulation, the authors have provided a scalable, memory-efficient solution to the NEGF bottlenecks.

While the current work assumes ballistic transport, the authors suggest that the framework can be extended to include scattering effects. Future research will also explore low-rank tensor approximations to further mitigate memory overhead and the integration of adaptive integration rules to enhance the efficiency of the MOR scheme.

Summary updated

BibTeX
@article{chen2015approximate,
  author = {Quan Chen and Jun Li and Chiyung Yam and Yu Zhang and Ngai Wong and Guanhua Chen},
  title = {{An approximate framework for quantum transport calculation with model order reduction}},
  journal = {Journal of Computational Physics},
  year = {2015},
  url = {https://openreview.net/forum?id=Quf9PqhqOO},
}
Figure